基於線上學習之視覺追蹤

機構典藏 > College of Engineering > Graduate Institute & Department of Computer Science and Information Engineering > Thesis > Item 987654321/102578

Please use this identifier to cite or link to this item: https://tkuir.lib.tku.edu.tw/dspace/handle/987654321/102578

Title:	基於線上學習之視覺追蹤
Other Titles:	An online learning based vision tracking
Authors:	林孝宗;Lin, Hsiao-Tsung
Contributors:	淡江大學資訊工程學系資訊網路與通訊碩士班顏淑惠;Yen, Shwu-Huey
Keywords:	線上學習;Haar like特徵;弱分類器;強分類器;追蹤;On-line boost;Haar-like Feature;Weak classifier;Strong classifier;Mean-Shift;Tracking
Date:	2014
Issue Date:	2015-05-04 09:59:18 (UTC+8)
Abstract:	即時的追蹤在電腦視覺中一直以來都是很重要的研究題目，本文裡我們提出了一個改善 Grabner和 Bischof於2006年所發表的On-line Boosting的追蹤方法。我們僅使用了簡單的Haar-like特徵來描述目標，並做篩選將表現比較好的特徵加入feature pool，而之後的強分類器訓練只需考慮這些特徵以有效的減少計算量。為了能適應前景/背景的變化，本文採取sample pool的方法進行分類器的即時更新，所賴以更新的訓練樣本則依照計算出的信任值，信任值高時更新sample pool的正樣本反之則更新負樣本；同時每隔一段時間，feature pool中的弱分類器的門檻值也會重新設定。我們並且使用background subtraction與Kalman filter來避免相似背景與雜訊的影響，並進行目標被遮蔽之後的路徑預測。實驗結果顯示，我們所提出的方法比原先的方法，能保有穩定且令人滿意的結果。 On-line tracking is main topic in computer vision. In this paper, we proposed a method based on On-line Boosting proposed by Grabner and Bischof. We only use Haar –like feature to describe target. We add good features into feature pool and then when training the strong classifier, we only use those features to decrease the computation. To adapt the changing of foreground and background we use sample pool to online update classifiers. When the confidence values are high, we update the positive samples in sample pool and vice versa. Meanwhile, we reset the threshold of weak classifiers every ten frames. We use background subtraction to avoid the effect of similar background and noise and Kalman filter to predict the object trajectory when it was occluded. From the experimental results, our method is more stable and accurate.
Appears in Collections:	[Graduate Institute & Department of Computer Science and Information Engineering] Thesis

Files in This Item:

File	Size	Format
index.html	0Kb	HTML	147	View/Open

Loading...